Download Document 8265714

Reviews • DRUG DISCOVERY TODAY: BIOSILICO
REVIEWS
DDT • Volume 10, Number 7 • April 2005
Computer prediction of drug
resistance mutations in proteins
Zhi Wei Cao,Lian Yi Han,Chan Juan Zheng,Zhi Lang Ji,Xin Chen,
Hong Huang Lin and Yu Zong Chen
Drug resistance is of increasing concern in the treatment of infectious diseases and
cancer. Mutation in drug-interacting disease proteins is one of the primary causes for
resistance particularly against anti-infectious drugs. Prediction of resistance
mutations in these proteins is valuable both for the molecular dissection of drug
resistance mechanisms and for predicting features that guide the design of new
agents to counter resistant strains. Several protein structure- and sequence-based
computer methods have been explored for mechanistic study and prediction of
resistance mutations. These methods and their usefulness are reviewed here.
Zhi Wei Cao
Lian Yi Han
Chan Juan Zheng
Xin Chen
Hong Huang Lin
Yu Zong Chen
Bioinformatics and Drug
Design Group,
Department of
Computational Science,
National University of
Singapore,
BLK SOC 1,
level 7,
Singapore 117543
*e-mail: [email protected]
Zhi Lang Ji
Department of Biology,
School of Life Sciences,
Xiamen University,
Xiamen 361000,
Fujian Province,
PR China
Drug resistance is a global public health problem
particularly for the treatment of infectious diseases
[1,2] and cancer [3]. It has become increasingly possible for cross country transmission of drug-resistant
organisms. There is an urgent need to develop resistance-evading drugs. Several mechanisms are responsible for drug resistance [4–8]. For infectious diseases,
resistance is primarily mediated by mutations in the
genes of infectious organisms that alter a drug’s
interaction with its corresponding target protein [4,5].
Considerable effort has been directed at the use
of computational methods for studying the molecular
mechanism of mutation-induced drug resistance
and for developing predictive tools. The availability
of the three-dimensional structure of drug targets
involved in disease enables the use of molecular
modeling and other structure-based approaches for
evaluation of structural features, molecular interactions, solvation and dynamical properties of drugprotein binding and their correlation to resistance
mutations [9–13]. Structure-derived binding energies
[9,14–16] and binding-site volume-based fitness
models [17] have also been used for facilitating the
prediction of resistance mutations.
Structural information is available for a relatively
small percentage of proteins. Thus methods for
predicting resistance mutations directly from sequences are highly useful and are being developed.
Interpretation programs have emerged for identifying and estimating the level of resistance [18,19].
Statistical learning methods such as neural networks
[20,21], support vector machines (SVM) [22] and
decision tree [23] have also shown promising potential for predicting resistance mutations.
Molecular modeling of drug resistance
mutations
Structural analysis of proteins that contain resistance
mutations indicated that mutations at drug-binding
sites usually alter the tight packing between the
binding drug and its receptor without substantial
change in overall conformation [4,24]. Comparison
of mutant and wild-type drug-receptor structures
showed that, in the majority of cases, the only
apparent change is in the pattern of local contact
and hydrogen bonding at a mutation site [24,25].
Hydrophobic effects have been found to be important
in several cases [10], but variation in local packing
1359-6446/04/$ – see front matter ©2005 Elsevier Ltd. All rights reserved. PII: S1359-6446(04)03377-5
www.drugdiscoverytoday.com
521
DDT • Volume 10, Number 7 • April 2005
REVIEWS
Reviews • DRUG DISCOVERY TODAY: BIOSILICO
interactions appears to be a major factor for the reduced
drug binding affinity leading to resistance [4]. Thus molecular mechanics, molecular dynamics simulation, and Monte
Carlo simulation methods are expected to be useful for
structural optimization and structure-based energetic
analysis of these mutations.
Molecular mechanics methods use atom–atom interaction energies for structural optimization and energetic
analysis of a ligand–protein complex [9,14,15]. Bonded
interactions are modeled by bond stretch, angle bending,
and torsion terms. Non-bonded interactions are modeled
by van der Waals and electrostatic terms. Hydrogen bonds
can either be modeled by a separate term [14] or they can
be included in van der Waals and electrostatic terms
[9,15]. Moreover, solvation and entropic effects may be
considered by using a simple solvation free energy model
[14,26] and side-chain entropy model [15] respectively.
Structural optimization involves the selection of low
energy molecular structures. One approach is to search
molecular conformations for identifying the structure of
the lowest energy. The other is energy minimization such
that molecular structure is varied towards the direction
of lower energy until it reaches the local minimum energy
configuration. In most cases, ligand-protein binding is
determined by non-bonded, hydrogen bond, solvation
and side-chain torsion interactions. Thus ligand-binding affinities are frequently estimated by using these
terms [9,14,15].
Molecular dynamics simulation methods derive trajectories of atomic positions of molecular motions and
dynamical fluctuations by solving Newton’s equations
governed by the same sets of bonded and non-bonded interactions used in molecular mechanics methods. Solvation
effect is described by using either explicit water molecules
or continuous medium models. Entropic effect is derived
from statistical analysis of trajectories of atomic positions.
These methods are useful for structural optimization,
motional and energetic analysis, and free energy computation [10–13].
The free energy difference between wild-type and mutant
systems, which is useful for indicating drug resistance
mutations, can be derived by using Monte Carlo simulation
[27] as well as molecular dynamics simulation [10–13].
Monte Carlo simulation methods generate a series of
molecular structures that are randomly distributed in the
molecular conformational space and conform to a certain
distribution pattern governed by the laws of statistical
mechanics. These randomly generated structures can then
be used to derive free energy difference between the wildtype and mutant structure by using the free energy perturbation method [27].
The structure of a ligand-protein complex and its mutants
often needs to be modeled from a template. Such a template
is usually the structure of a different mutational variant of
the same protein or its complex with a different ligand
[9–16]. The modeled structures of both wild-type and
522
www.drugdiscoverytoday.com
mutant HIV-1 protease–inhibitor complexes have been
found to be consistent with the crystallographic structures,
with root mean square differences ranging from 0.5A to
1.2A [9,14,15], which suggests that the quality of these
modeled structures reaches the level useful for facilitating
structure-based study and prediction of resistance mutations.
The ligand-protein binding energies computed from
molecular mechanics and the free energies computed
from molecular dynamics for several wild-type and mutant
ligand–protein complexes showed significant correlation
with the observed binding affinities [9–16,26], indicating
that these energy functions are useful for facilitating energetic analysis and the prediction of resistance mutations.
Structure-based prediction of drug resistance
mutations
Structure-base virtual screening has been widely used for
designing new drugs [28,29]. To achieve high-speed screening of a large number of compounds, efficient computational
procedures have been routinely applied to structural
optimization and scoring of docked ligand-protein structures [30–34]. Some of these procedures are very similar
to those used for the molecular study of drug resistance
mutations [9,13–16], which raises the possibility of using
virtual screening approaches to indicate possible drug
resistance mutations [14]. While more accurate in modeling
resistance mutations than those used in virtual screening
studies [10–13], a full molecular dynamics procedure has
yet to be employed in a general virtual screening process
partly owing to its computationally intensive nature.
Thus, procedures that either use molecular mechanics
alone or molecular mechanics plus a small run of molecular dynamics for structural optimization has been the
primary choice for structure-based prediction of resistance
mutations in a virtual screening process [9,14].
Structural models and energy functions
Most models of mutant drug–protein complexes are based
on a starting crystal structure of the wild-type protein
complexed with the same drug [9,14,15]. Each mutation
is introduced by stripping the amino acid down to the
atom and replacing it by the side chain of the new amino
acid. In some studies, the atom is also kept intact for
mutations between amino acids R, K and Q, as these are
relatively large amino acids and normally located at the
protein surface [9]. The mutant structure is then optimized by conformation search for the local residues to
release structural clash among them and the binding drug,
which is conducted by variation of rotatable bonds of the
drug and those of the side-chain of the amino acids in
contact with the drug. This is followed by energy minimization to allow the mutant structure to find the local
minimum energy conformation [9,15]. In some studies,
molecular dynamics simulation is conducted to allow the
mutant structure to reach a more appropriate local minimum energy conformation [14]. The procedure of this
modeling process and the subsequent prediction of resistance mutations from the derived mutant three-dimensional structure is illustrated in Figure 1.
Energy minimization is typically conducted for no less
than a few hundred iterations, compared to the <300 steps
of energy minimization used in the majority of virtual
screening studies [34]. Molecular dynamics simulation is
typically performed for picoseconds and at a temperature
of 300K. The drug-protein interaction energy is computed
by the following empirical potential energy function or
its variations that includes non-bonded van der Waals and
electrostatic interactions, atomic solvation free energy
terms, side chain conformational entropy, and hydrogen
bond energy (Equation 1)
 Aij Bij 

 qi q j
V = ∑ vdw  12 − 6  + ∑electr 
+ E( rij )
rij 
 rij

 ε r rij
+ ∑ atoms i ∆σ i Ai + ∑ residue k α k N k + ∑ H bondsV H
Aij and Bij are van der Waals parameters; εr is the dielectric constant, qi and qj are the partial charges of the i-th and
j-th atoms, and rij is the distance between them; E(rij) is
the electrostatic interaction correction term used in some
studies [9], ∆σi is the atomic solvation parameter and is
the solvent-accessible surface area of the i-th atom [35], Nk
is the number of rotatable bonds and αk the coefficient of
the k-th residue [36], and VH is the hydrogen-bond energy
[14]. Various sets of potential energy parameters, which
are called force fields, have been used for the prediction
of resistance mutations, which includes AMBER [14],
CHARMM [26], UFF [9], ECEPP/3 [15] and CVFF [16].
Apart from the molecular modeling approach, two
other structure-based methods have also been used for
predicting resistance mutations. One is the evolutionary
simulation model that analyses resistance mutations that
produces viable viruses using Michaelis-Menten kinetics
derived from a simple measure of volume complementarity between a mutant and its binding drug [17]. The other
is a neural network model that uses structural features
generated from homology modeled mutant structures for
classification of resistance mutations [21].
Molecular modeling of HIV-1 protease consistently
showed a good correlation with experimentally estimated
binding affinities, indicating the usefulness of the molecular modeling approach for prediction of resistance mutations.
The prediction accuracy for HIV-1 reverse transcriptase is
substantially lower that those of HIV-1 protease. Significant
conformational flexibility in this enzyme [39], which is
inadequately modeled by molecular modeling methods,
is likely a factor for the lower accuracy.
Several other factors have not been adequately described
by molecular modeling methods. These include interactions
with water, hydrophobic effect and electrostatic interactions
involving aromatic rings, which likely lead to a certain
degree of error in ligand-protein structural optimization
and the scoring of resistance mutations. Proper modeling
of these interactions and further refinement of currently
3D structure of wild-type protein–drug complex
At each mutation site, strip the amino acid down
to the Cγ atom for R, K, Q
to the Cβ atom for other amino acids
Add new amino acid side-chain to the Cγ or Cβ atom
Conduct conformation search for the binding
drug and side-chain of the mutated and neighboring
amino acids to release bad contact between them
Conduct energy minimization for the whole
complex to find the local minimum energy conformation
No
Yes
Further
optimization?
Conduct molecular dynamics simulation
for the whole complex to find
more appropriate local
minimum energy conformation
3D structure of mutant protein–drug complex
Prediction performance
Table 1 summarizes the results of structure-based prediction studies of resistance mutations in HIV-1 protease
[9,13–16,37] and HIV-1 reverse transcriptase [14,27,38].
The performance of these studies is primarily measured
by the correlation coefficient, the R value, between
changes in the computed binding energy and those in the
experimentally estimated binding affinities. Prediction
accuracy for resistance and non-resistance mutations, the
P value, has also been used in some studies [13,17,21].
The computed R values range from 0.31 to 0.90 [9,14–16].
The computed P values are in the range of 57%–86% [13]
for the molecular modeling methods and 45%–70% for
the other two structure-based methods [17,21].
Drug binding energy in mutant complex
larger than that in wild-type complex
by a specified amount?
Yes
Drug resistance mutations
No
Non-resistance mutations
Drug Discovery Today
FIGURE 1
Schematic diagram illustrating the process of molecular modeling
of the three-dimensional structure of a mutant protein–drug
complex from that of the wild-type protein–drug complex, and the
prediction of drug resistance mutations from the derived mutant
three-dimensional structure.
www.drugdiscoverytoday.com
523
Reviews • DRUG DISCOVERY TODAY: BIOSILICO
REVIEWS
DDT • Volume 10, Number 7 • April 2005
DDT • Volume 10, Number 7 • April 2005
REVIEWS
TABLE 1
Structure-based prediction of drug resistance mutations
Reviews • DRUG DISCOVERY TODAY: BIOSILICO
Method and year of report
Protein
Drugs
Number of mutantdrug complexes
Molecular modeling 1999 [9]
HIV-1 protease Indinavir, Saquinavir
17
R = 0.68–0.76
Molecular modeling 2001 [13]
HIV-1 protease Amprenavir, Indinavir, Nelfinavir,
Ritonavir, Saquinavir
33
P = 57%–86%
22
R = 0.64
Molecular modeling with simple
HIV-1 protease MK639, Saquinavir, SB203386,
optimization scoring procedure 2001 [14]
U89360E, VX478
Reported prediction
accuracy (R or P value)
Molecular modeling 2002 [16]
HIV-1 protease Ritonavir
12
R = 0.7
Fitness evolution model 2002 [17]
HIV-1 protease Indinavir, Nelfinavir, Ritonavir,
Saquinavir
11
P = 45% (81% for
resistance site prediction)
Molecular modeling 2003 [15]
HIV-1 protease Amprenavir, Indinavir, Lopinavir,
Nelfinavir, Ritonavir, Saquinavir
750
R = 0.87 for 360
complexes.
P = 86% and 92%
Neural network model of homology
modeled structures 2003 [21]
HIV-1 protease Indinavir
38
P = 60%–70%
Molecular Dynamics simulation 2004 [37] HIV-1 protease Indinavir
12
R = 0.98
Molecular modeling 1997 [38]
HIV-1 reverse
transcriptase
8-CL TIBO, α-APA
8
R = 0.80–0.97
Monte Carlo Simulation 2000 [27]
HIV-1 reverse
transcriptase
8-CL TIBO
2
R = 0.78
Nevirapine, TIBO R82913
13
R = 0.31
Molecular modeling with simple
HIV-1 reverse
optimization scoring procedure 2001 [14] transcriptase
Reported prediction accuracy is given by the R value (correlation coefficient between changes in the computed binding energy and those in the experimentally
estimated binding affinities) or the P value (percentage of correctly predicted resistance and non-resistance mutations).
used optimization/scoring force field parameters may help
improve the quality of modeled mutant structure and the
accuracy of the scoring functions.
anti-HIV drugs. The highest scored degree of resistance is
used for assigning one of the three resistance levels (no
evidence of resistance, possible resistance, and resistance)
for each drug [19].
Sequence-based prediction of resistance mutations
Sequence-based methods predict resistance mutations by
using the rules or classifiers derived from statistical analysis
of the sequences of resistant and non-resistant samples. A
straightforward approach is to identify known resistance
mutations directly from the sequence, which is similar to
the motif approach for protein functional analysis [40].
Genotype interpretation systems, such as Stanford HIVdb
(HIVdb or HIV-SEQ), AntiRetroScan (ARS), and Visible
Genetics/Bayer Diagnostics (VGI), have been developed
for estimating the level of resistance mutations in HIV-1
protease and transcriptase based on this approach [18,19].
HIVdb scans a sequence for known resistance mutations and generates scores for each of the 16 anti-HIV
drugs by computing the sum of the penalties retrieved
from a table of resistance mutations of defined values of
resistance level for each drug and its target protein (HIV-1
protease or reverse transcriptase). From the final score, one
of the five susceptibility levels (susceptible, potential lowlevel resistance, low-level resistance, intermediate resistance, and high-level resistance) is assigned for each drug
[18,19]. ARS is similar to HIVdb and uses the same penalty
functions plus an additional set of 8–12 rules to determine
the final scores [19]. VGI extracts known resistance mutations and compares them against a set of 85 rules for
determining the level of resistance for each of the 16
524
www.drugdiscoverytoday.com
Statistical learning methods
Statistical learning methods such as neural networks (NN)
[20,21], support vector machines (SVM) [22] and decision
tree (DT) [23] have been used for developing classifiers
and rules for predicting resistance mutations. In these
approaches, protein sequence is represented by a feature
vector xi, with 20 binary bits (one for each type of amino
acid) for each sequence position as its components.
Samples of resistance and non-resistance mutations are
used for training a statistical learning system to derive the
classifiers and rules of resistance mutations [20–22]. These
classifiers and rules can then be used to predict resistance
mutations from the sequence of a mutant protein. The
procedure for such a prediction is illustrated in Figure 2
where an example of SVM prediction of resistance mutations is provided.
NN trains a hidden-layer-containing network and uses
its outcomes for pattern recognition and classification
of the input feature vectors [20,21]. A classifier for NN is
y = g Σj w0j hj, where w0j is the output weight of a hidden
node j to an output node, g is the output function, hj is
the value of a hidden layer node: hj = σ (Σj wji xj + wj), wji
is the input weight from an input node i to a hidden node
j, wj is the threshold weight from an input node of value
1 to a hidden node j, and σ is a sigmoid function. Known
resistance and non-resistance samples are used for training
a NN such that all the weights are determined, and the
resulting classifier can be used for determining whether
or not a new input sequence is resistant to a particular
drug [20].
SVM constructs a hyperplane in a hyperspace for separating two groups of feature vectors with a maximum
margin [22]. A classifier for a linear SVM is yi = (w. xj + b),
where w is a vector normal to the hyperplane, is the perpendicular distance from the hyperplane to the origin
with as the Euclidean norm of w, yi = +1 for resistance
and yi = –1 for non-resistance mutations. The distribution
functions of SVM can also be used for generating regression models [22]. Known resistance and non-resistance
samples are used to train a SVM such that the hyperspace,
distribution functions, and the parameters are determined. The resulting classifier can be used for predicting
resistance and non-resistance mutations based on the value
of yi, and the trained regression model used for indicating the level of resistance from the regression model.
DT is an acyclic graph in which its interior vertices specify testing of a single attribute (sequence position) of a
feature vector and its leaves indicate the classes of the
attribute (resistance or non-resistance mutation) [23]. A
Sequence of mutant protein
Translate sequence into a specified representation such as a
binary code (0001; 0010; 0101; …) to form a feature vector
Feature vector
Map feature vector
to a multi-dimensional
input space
Z
A = (1, 1, 1)
B = (0, 1, 1)
E = (0, 0, 0)
F = (1, 0, 1)
F
B
E A
X
Project feature vector to a
higher dimensional
feature space where a
hyper-plane separates
resistance and
non-resistance mutants
Input
space
Input space
Y
Feature space
Resistance
mutants
Resistance
mutants
Border
New
border
Non-resistance
mutants
Non-resistance
mutants
Feature vector on
resistance mutant side
of hyper-plane?
Yes
Drug resistance mutations
No
Non-resistance mutations
classifier of DT is constructed by recursively splitting the
sample set, with each subset giving rise to one new vertex
connected with an edge to its parent. This procedure continues until all samples at each leaf belong to the same
class. Classification of a sample is achieved by running
through the tree from the root to a leaf according to the
values (amino acids) of the attributes of the protein sequence
that appear on this path [23].
Prediction performance
Table 2 summarizes the results of sequence-based prediction studies of resistance mutations in HIV-1 protease and
HIV-1 reverse transcriptase [20–23,41]. The performance
of statistical learning method has been measured by the
overall prediction accuracy for resistance and non-resistance
mutations P, prediction accuracy for resistance mutations
Pr, and prediction accuracy for non-resistance mutations
Pn. The R value has been used in one study [22]. The number of samples in most of these studies is significantly
higher than those in structure-based methods, reflecting
the more extensive application range of sequence-based
methods.
The computed Pr and Pn values are in the range of
58%–97% and 62%–97% respectively, and these are >90%
in majority of the studies. The computed R values are in
the range of 0.78–0.89 for HIV-1 protease, which are comparable to those obtained by structure-based methods,
and 0.54–0.85 for HIV-1 reverse transcriptase, which are
slightly better than those of structure-based methods. These
results suggest that sequence-based methods are capable of
equally accurate prediction of resistance mutations. As in
the case of structure-based methods, the accuracies for
HIV-1 reverse transcriptase appear to be substantially lower
than those for HIV-1 protease. Conformational flexibility
in HIV-1 reverse transcriptase [39], which is insufficiently
represented in the feature vectors of statistical learning
methods, is likely to be a major factor for the reduced
accuracies.
The performance of sequence-based methods depends
on the proper representation of mutant sequences as well
as adequate training using a sufficiently diverse set of samples. While useful for distinguishing different mutations,
it has been pointed out that the binary representation of
amino acids used in most studies lacks biological basis
and unnecessarily expands the input space such that it
complicates the statistical learning task [42]. Thus, a more
appropriate representation of amino acids is useful for further improving the performance of the sequence-based
methods.
Drug Discovery Today
FIGURE 2
Schematic diagram illustrating the process of the prediction of resistance
mutations from the sequence of a mutant by using a statistical learning method support vector machines. A,B: feature vectors of non-resistant mutants; E,F: feature
vectors of resistant mutants; black filled squares, non-resistant mutants; green circles,
resistant mutants.
Prediction of resistance mutations by using
simple rules
So far, both structure- and sequence-based methods for
predicting resistance mutations have been primarily developed and tested on two proteins, HIV-1 protease and HIV-1
reverse transcriptase. This is because of the availability of
www.drugdiscoverytoday.com
525
Reviews • DRUG DISCOVERY TODAY: BIOSILICO
REVIEWS
DDT • Volume 10, Number 7 • April 2005
DDT • Volume 10, Number 7 • April 2005
REVIEWS
TABLE 2
Reviews • DRUG DISCOVERY TODAY: BIOSILICO
Sequence-based prediction of drug resistance mutations
Method and year of
report
Protein
Drugs
Number of mutant-drug Reported prediction
accuracy (P, Pr, Pn or R
samples (N, Nr, Nn)
value)
HIVdb (HIV SEQ) 2002 [41] HIV-1 protease
Amprenavir, Indinavir, Nelfinavir, Ritonavir,
Saquinavir
N = 529
P = 79%–97%
HIVdb 2004 [19]
HIV-1 protease
Amprenavir, Indinavir, Nelfinavir, Ritonavir,
Saquinavir
N = 2460
Pr = 81%–95%
Pn = 96%–97%
ARS 2004 [19]
HIV-1 protease
Amprenavir, Indinavir, Nelfinavir, Ritonavir,
Saquinavir
N = 2460
VGI 2004 [19]
HIV-1 protease
Amprenavir, Indinavir, Nelfinavir, Ritonavir,
Saquinavir
N = 2460
Decision Tree 2002 [23]
HIV-1 protease
Amprenavir, Indinavir, Nelfinavir, Ritonavir,
Saquinavir
N = 2148
Nr = 32
Nn = 811
Neural Network 2003 [21]
HIV-1 protease
Saquinavir
Neural Network 2003 [20]
HIV-1 protease
Lopinavir
SVM (Geno2pheno) 2003
[22]
HIV-1 protease
Amprenavir, Atazanavir, Indinavir, Lopinavir,
Nelfinavir, Ritonavir, Saquinavir
Pr = 92%
Pn = 93%
N = 3683
R = 0.78–0.89
HIV-1 reverse transcriptase
AZT, abacavir, didanosine, lamivudine,
stavudine, zalcitabine, delavirdine, efavirenz,
nevirapine, tenofovir
N = 4780
ARS 2004 [19]
HIV-1 reverse transcriptase
AZT, abacavir, didanosine, lamivudine,
stavudine, zalcitabine, delavirdine, efavirenz,
nevirapine, tenofovir
N = 4780
AZT, abacavir, didanosine, lamivudine,
stavudine, zalcitabine, delavirdine, efavirenz,
nevirapine, tenofovir
N = 4780
AZT, abacavir, didanosine, lamivudine,
stavudine, zalcitabine, delavirdine, efavirenz,
nevirapine
N = 4104
AZT, abacavir, didanosine, lamivudine,
stavudine, zalcitabine, delavirdine, efavirenz,
nevirapine, tenofovir
N = 6018
SVM 2003 [22]
HIV-1 reverse transcriptase
HIV-1 reverse transcriptase
P = 78%
Nr = 267
HIVdb 2004 [19]
Decision Tree 2002 [23]
Pr = 82%–90%
Nn = 1055
Zidovudine, didanosine, zalcitabine, stavudine, N=954
abacavir, lamivudine, delavirdine, efavirenz,
nevirapine
HIV-1 reverse transcriptase
Pr = 84%–97%
Pn = 94%–97%
Pn = 87%–90%
HIVdb (HIV SEQ) 2002 [41] HIV-1 reverse transcriptase
VGI 2004 [19]
Pr = 87%–97%
Pn = 94%–97%
P = 53%–96%
Pr = 90%
Pn = 97%
Pr = 95%
Pn = 97%
Pr = 93%
Pn = 96%
Pr = 58%–92%
Pn = 62%–97%
R = 0.54–0.85
N, Nr, and Nn are the number of all, resistant, and non-resistant mutant-drug samples, respectively. P, P r, Pn values are the percentage of correct predictions for all mutations, resistance
mutations, and non-resistance mutations, respectively.
a larger amount of drug resistance mutation data and a
higher number of high-resolution three-dimensional
structures for these two proteins. For instance, comprehensive resistance data for these two proteins are collected
in the HIVdb database [18]. There are 198 and 102 entries
of the three-dimensional structure of HIV-1 protease- and
reverse transcriptase-inhibitor complexes respectively in
the PDB database [43]. Significantly less data and structural information is available for other proteins with
known resistance mutations, which likely contributes to
the lack of progress in the development and testing of
computer prediction methods for these proteins.
Known resistance mutations in other proteins have
been used in combination with experimental molecular
techniques as markers for predicting drug resistance
mutations [44–49]. These exploit the observation that spe526
www.drugdiscoverytoday.com
cific mutations found in resistance strains are absent in
susceptible strains. Based on the analysis of clinical samples
and the results of mutagenesis, simple rules ranging from
identification of a single point mutation to more complex mutation selection algorithm can be derived for the
prediction of drug resistance [45]. Experimental techniques
capable of differentiating between a single wild-type sequence
and mutant sequences can then be used to detect these
mutations and predict drug resistance strains [46].
The simple rules generated and applied in these experimental studies can be directly used for developing a
computer prediction system in a similar fashion like the
HIV resistant genotype interpretation systems HIVdb,
ARS, and VGI [18,19]. To predict mutations of a protein
resistant to a particular drug, the sequence of that protein
can be scanned to identify mutations that match the simple
REVIEWS
DDT • Volume 10, Number 7 • April 2005
TABLE 3
Protein
Drugs
Rule and Year of Report
Number of Prediction Accuracy
Samples
(P, Pr, Pn value)
(N, Nr, Nn)
Mycobacterium tuberculosis arabinosyltransferase B
Ethambutol
embB codon 306 mutation, 1997 [46,64]
N = 118
Mycobacterium tuberculosis catalase-peroxidase
Isoniazid
katG codon 315 mutation, 2000 [46,65]
N = 79
Pr = 33%~45%
Mycobacterium tuberculosis β subunit of RNA
polymerase
Rifampicin
Any mutation at rpoB codon 531/526/516,
2002 [46,66,67]
N = 20
P = 90%~98%
Plasmodium falciparum multi-drug resistance protein 1 Choloroquine
Any mutation at fmdr1 codon 86/1042/1246, Nr = 40
2000 [68]
Nn = 17
Pr = 60%~68%
Pr = 86%~100%
Plasmodium falciparum dihydrofolate reductase
SulfadoxineCodon 59 mutation, 2003 [47]
pyrimethamine
N = 327
Pr = 81%
Hepatitis C INF-sensitive-determining region of
nonstructural 5A protein
Interferon
Nr = 36
Pr = 58%
Nn = 14
Pn = 86%
Helicobacter pylori 23S rRNA gene
Clarithromycin A2142G or A2143G mutation, 2001 [44]
N = 299
Pr =42%
Pr = 51%
Amino acid 2218 to be H, 2001 [69]
Neisseria Meningitidis penicillin-binding protein 2
Penicillin
I566V mutation, 2003 [49]
Nr = 30
HIV-1 reverse transcriptase
Abacavir
L74V or NRTI MDR or any 3 mutations of
41/184/210/215, 2002 [45]
N = 307
Codon 1612 Adenine deletion, 2004 [70]
N = 148
Pr = 45%
Nr = 9
Pr = 100%
Human matrix metalloproteinase 3
5-fluorouracilcisplatin
PKC412,
SU5614,
K-252a,
D-64406,
Human receptor tyrosine kinase FLT3
D-65476,
G697R mutation, 2004 [48]
DQPPC,
AGL2043,
TMPPP,
GTP-14564
(anti-leukemia)
rules associated with that protein and the corresponding
drug. Table 3 gives the protein-drug systems that have
known drug resistance mutations and sufficiently accurate
prediction rules reported in the literature. The simple rules
for these protein-drug systems and the reported prediction
accuracy derived from these rules are also given in Table 3.
Based on the test of the currently available samples, these
rules are capable of predicting resistance and non-resistance mutations at accuracies of 42%–100% and 77%–86%
respectively, which are compared to those of 81%–97%
and 91%–97% from the HIV resistant genotype interpretation systems HIVdb, ARS and VGI [18,19]. This suggests
that these simple rules have a certain capacity for facilitating the prediction of specific drug resistance mutations
and they may be used as the basis for developing more
sophisticated interpretation systems like those of HIVdb,
ARS and VGI [18,19].
Conclusions and perspectives
Both structure-based and sequence-based methods consistently show a promising capability for predicting resistance
mutations. Structure-based methods are particularly useful
Reviews • DRUG DISCOVERY TODAY: BIOSILICO
Prediction of drug resistance mutations by using simple rules
Pr = 88%
Pn = 77%
for mechanistic study and the prediction of resistant protein
variants with little or no preliminary knowledge of known
resistance mutations. However, they depend on the availability of a structural template, which significantly limits
their application range. Advances in structural genomics
are expected to expand the application range of structurebased methods to more extensive sets of disease proteins
and drugs. Inadequately described interactions need to be
more properly modeled. For instance, algorithms for modeling main-chain and side-chain conformational flexibility
[50], ligand-protein-solvent interactions [51], hydration
effects [52] and electrostatic interactions involving aromatic
rings [53] can be incorporated into structure-based methods
to increase their prediction accuracies.
No structural template is required by sequence-based
methods. However, a sufficiently diverse set of resistance
and non-resistance samples is needed for training a statistical learning system. Thus these methods are not applicable for disease proteins and drugs with little or no resistance
mutation data. Mining of the resistance and non-resistance
mutation data from the literature [4,5,54,55] and other
sources [56] is a key to more extensive exploration of
www.drugdiscoverytoday.com
527
DDT • Volume 10, Number 7 • April 2005
REVIEWS
Reviews • DRUG DISCOVERY TODAY: BIOSILICO
sequence-based and rule-based methods. Databases such
as HIVdb [18] and a rational database of HIV protein
sequences of AIDS patients [57] that provide resistance
mutation data are useful resources for serving this purpose,
and more such databases are desired. Studies of resistance
mutation patterns [44–46,58] are also important for determining the phenomenon and mechanism of drug resistance.
Moreover, the performance of sequence-based methods
may be further improved by introducing a more biologically
meaningful representation of amino acids [42].
Sequence-based methods are less effective in prediction
of resistant mutations for new drugs, when only limited
training datasets are available. By contrast, structure-based
methods are capable of predicting mutations resistant to
a new drug, if a high quality template structure of the
drug-protein complex is available. Moreover, sequencebased methods are usually not intended for facilitating
mechanistic study of resistance mutations. Some of these
methods, such as DT [23,59,60], NN regression [61] and
SVM regression [22,62,63], have the capacity for providing
the contribution of specific attributes (structural and
physicochemical properties of amino acids) to a classification (resistance or non-resistance mutation). By introducing
a more biologically meaningful representation of amino
acids [42], this capacity may be explored for probing structural and physicochemical features contributing to resistance
mutations.
Acknowledgements
The authors would like to thank the support from the
EU-China Scientific Collaborative Project Shanghai
Commission for Science and Technology (03DJ14011),
China ‘863’ National High-Tech Program (2003AA231010,
2004BA711A21), China ‘973’ National Key Basic Research
Program (2001CB510203, 2003CB715901), and Singapore
NUS-ARF (R-151–000–031–112). Zhi Wei Cao and Yu Zong
Chen are also affiliated with the Shanghai Center for
Bioinformatics Technology.
References
1 Chaix-Couturier, C. et al. (2000) HIV-1 drug
resistance genotyping. A review of clinical and
economic issues. Pharmacoeconomics 18, 425–433
2 Liu, H.H. (1999) Antibiotic resistance in
bacteria. A current and future problem. Adv.
Exp. Med. Biol. 455, 387–396
3 Lehne, G. et al. (1998) Challenging drug
resistance in cancer therapy–review of the First
Nordic Conference on Chemoresistance in
Cancer Treatment, October 9th and 10th, 1997.
Acta Oncol 37 (5), 431-439
4 Erickson, J.W. and Burt, S.K. (1996) Structural
mechanisms of HIV drug resistance. Annu. Rev.
Pharmacol. Toxicol. 36, 545–571
5 Blanchard, J.S. (1996) Molecular mechanisms of
drug resistance in Mycobacterium tuberculosis.
Annu. Rev. Biochem. 65, 215–239
6 Borst, P. (1991) Genetic mechanisms of drug
resistance. A review. Acta Oncol. 30, 87–105
7 Remy, S. et al. (2003) A novel mechanism
underlying drug resistance in chronic epilepsy.
Ann. Neurol. 53, 469–479
8 Schafer, J.M. et al. (2002) A mechanism of drug
resistance to tamoxifen in breast cancer. J.
Steroid Biochem. Mol. Biol. 83, 75–83
9 Weber, I.T. and Harrison, R.W. (1999) Molecular
mechanics analysis of drug-resistant mutants of
HIV protease. Protein Eng. 12, 469–474
10 Sussman, F. et al. (1997) Solvation effects are
responsible for the reduced inhibitor affinity of
some HIV-1 PR mutants. Protein Sci. 6, 1024–1030
11 Rick, S.W. et al. (1998) Molecular mechanisms
of resistance: free energy calculations of
mutation effects on inhibitor binding to HIV-1
protease. Protein Sci. 7, 1750–1756
12 Piana, S. et al. (2002) Drug resistance in HIV-1
protease: Flexibility-assisted mechanism of
compensatory mutations. Protein Sci. 11,
2393–2402
13 Wang, W. and Kollman, P.A. (2001)
Computational study of protein specificity: the
molecular basis of HIV-1 protease drug resistance.
Proc. Natl. Acad. Sci. U. S. A. 98, 14937–14942
14 Chen, Y.Z. et al. (2001) Can an
optimization/scoring procedure in ligandprotein docking be employed to probe drug-
528
www.drugdiscoverytoday.com
15
16
17
18
19
20
21
22
23
24
25
26
resistant mutations in proteins? J. Mol. Graph.
Model. 19, 560–570
Shenderovich, M.D. et al. (2003) Structure-based
phenotyping predicts HIV-1 protease inhibitor
resistance. Protein Sci. 12, 1706–1718
Nair, A.C. et al. (2002) Computational studies of
the resistance patterns of mutant HIV-1 aspartic
proteases towards ABT-538 (ritonavir) and
design of new derivatives. J. Mol. Graph. Model.
21, 171–179
Stoffler, D. et al. (2002) Evolutionary analysis of
HIV-1 protease inhibitors: Methods for design
of inhibitors that evade resistance. Proteins 48,
63–74
Rhee, S.Y. et al. (2003) Human
immunodeficiency virus reverse transcriptase
and protease sequence database. Nucleic Acids
Res. 31, 298–303
Zazzi, M. et al. (2004) Comparative evaluation
of three computerized algorithms for prediction
of antiretroviral susceptibility from HIV type 1
genotype. J. Antimicrob. Chemother. 53, 356–360
Wang, D. and Larder, B. (2003) Enhanced
prediction of lopinavir resistance from
genotype by use of artificial neural networks.
J. Infect. Dis. 188, 653–660
Draghici, S. and Potter, R.B. (2003) Predicting
HIV drug resistance with neural networks.
Bioinformatics 19, 98–107
Beerenwinkel, N. et al. (2003) Geno2pheno:
Estimating phenotypic drug resistance from
HIV-1 genotypes. Nucleic Acids Res. 31,
3850–3855
Beerenwinkel, N. et al. (2002) Diversity and
complexity of HIV-1 drug resistance: a
bioinformatics approach to predicting
phenotype from genotype. Proc. Natl. Acad. Sci.
U. S. A. 99, 8271–8276
Ala, P.J. et al. (1997) Molecular basis of HIV-1
protease drug resistance: structural analysis of
mutant proteases complexed with cyclic urea
inhibitors. Biochemistry 36, 1573–1580
Rozwarski, D.A. et al. (1998) Modification of the
NADH of the isoniazid target (InhA) from
Mycobacterium tuberculosis. Science 279, 98–102
Avram, S. et al. (2002) Comparative study of
27
28
29
30
31
32
33
34
35
36
37
38
some energetic and steric parameters of the wild
type and mutants HIV-1 protease: a way to
explain the viral resistance. J. Cell. Mol. Med. 6,
251–260
Smith, M.B. et al. (2000) Monte Carlo
calculations on HIV-1 reverse transcriptase
complexed with the non-nucleoside inhibitor
8-Cl TIBO: contribution of the L100I and
Y181C variants to protein stability and
biological activity. Protein Eng. 13, 413–421
Kuntz, I.D. (1992) Structure-based strategies for
drug design and discovery. Science 257, 1078–1082
Blundell, T.L. (1996) Structure-based drug
design. Nature 384 (Suppl. 6604), 23–26
Rarey, M. et al. (1996) A fast flexible docking
method using an incremental construction
algorithm. J. Mol. Biol. 261, 470–489
Jones, G. et al. (1997) Development and
validation of a genetic algorithm for flexible
docking. J. Mol. Biol. 267, 727–748
Wang, J. et al. (1999) Flexible ligand docking: a
multistep strategy approach. Proteins 36, 1–19
Charifson, P.S. et al. (1999) Consensus scoring:
A method for obtaining improved hit rates from
docking databases of three-dimensional
structures into proteins. J. Med. Chem. 42,
5100–5109
Muegge, I. and Martin, Y.C. (1999) A general
and fast scoring function for protein-ligand
interactions: a simplified potential approach.
J. Med. Chem. 42, 791–804
Eisenberg, D. and McLachlan, A.D. (1986)
Solvation energy in protein folding and
binding. Nature 319, 199–203
Wang, R. et al. (1998) SCORE: a new empirical
method for estimating the bind affinity of a
protein-ligand complex. J. Mol. Model. 4, 379–394
Chen, X. et al. (2004) Molecular dynamics
simulations of 14 HIV protease mutants in
complexes with indinavir. J Mol Model (Online)
10, 373–381
Kroeger Smith, M.B. et al. (1997) Molecular
modeling of HIV-1 reverse transcriptase drugresistant mutant strains: implications for the
mechanism of polymerase action. Protein Eng.
10, 1379–1383
39 Arnold, E. et al. (1996) Targeting HIV reverse
transcriptase for anti-AIDS drug design:
structural and biological considerations for
chemotherapeutic strategies. Drug Des Discov
13, 29–47
40 Hu, J.S. et al. (2002) A flexible approach of
measuring mRNA expression levels of any genes
using a universal oligonucleotide microarray.
FASEB JOURNAL 16, A543–A543
41 Puchhammer-Stockl, E. et al. (2002)
Comparison of virtual phenotype and HIV-SEQ
program (Stanford) interpretation for predicting
drug resistance of HIV strains. HIV Med. 3,
200–206
42 Thomson, R. et al. (2003) Characterizing
proteolytic cleavage site activity using bio-basis
function neural networks. Bioinformatics 19,
1741–1747
43 Berman, H.M. et al. (2000) The Protein Data
Bank. Nucleic Acids Res. 28, 235–242
44 van Doorn, L.J. et al. (2001) Accurate prediction
of macrolide resistance in Helicobacter pylori by a
PCR line probe assay for detection of mutations
in the 23S rRNA gene: multicenter validation
study. Antimicrob. Agents Chemother. 45,
1500–1504
45 Walter, H. et al. (2002) Prediction of abacavir
resistance from genotypic data: impact of
zidovudine and lamivudine resistance in vitro and
in vivo. Antimicrob. Agents Chemother. 46, 89–94
46 Victor, T.C. et al. (2002) Prediction of drug
resistance in M. tuberculosis: molecular
mechanisms, tools, and applications. IUBMB
Life 53, 231–237
47 Talisuna, A.O. et al. (2003) Population-based
validation of dihydrofolate reductase gene
mutations for the prediction of sulfadoxinepyrimethamine resistance in Uganda. Trans. R.
Soc. Trop. Med. Hyg. 97, 338–342
48 Cools, J. et al. (2004) Prediction of resistance to
small molecule FLT3 inhibitors: implications for
molecularly targeted therapy of acute leukemia.
Cancer Res. 64, 6385–6389
49 Stefanelli, P. et al. (2003) Prediction of decreased
susceptibility to penicillin of Neisseria
meningitidis strains by real-time PCR. J. Clin.
Microbiol. 41, 4666–4670
50 Zavodszky, M.I. et al. (2004) Modeling correlated
main-chain motions in proteins for flexible
molecular recognition. Proteins 57, 243–261
51 Floriano, W.B. et al. (2004) HierVLS hierarchical
docking protocol for virtual ligand screening of
large-molecule databases. J. Med. Chem. 47, 56–71
52 Liu, Y. and Beveridge, D.L. (2002) Exploratory
studies of ab initio protein structure prediction:
multiple copy simulated annealing, AMBER
energy functions, and a generalized
born/solvent accessibility solvation model.
Proteins 46, 128–146
53 Burghardt, T.P. et al. (2002) Cation-pi
interaction in a folded polypeptide. Biopolymers
63, 261–272
54 Cheung, P.K. et al. (2004) 2004: which HIV-1
drug resistance mutations are common in
clinical practice? AIDS Rev. 6, 107–116
55 Anderson, T.J. (2004) Mapping drug resistance
genes in Plasmodium falciparum by genomewide association. Curr. Drug Targets Infect.
Disord. 4, 65–78
56 Weinstein, J.N. et al. (1997) An informationintensive approach to the molecular
pharmacology of cancer. Science 275, 343–349
57 Tamalet, C. et al. (2003) Resistance of HIV-1 to
multiple antiretroviral drugs in France: a 6-year
survey (1997-2002) based on an analysis of over
7000 genotypes. AIDS 17, 2383–2388
58 Yahi, N. et al. (1999) Mutation patterns of the
reverse transcriptase and protease genes in
human immunodeficiency virus type 1-infected
patients undergoing combination therapy:
survey of 787 sequences. J. Clin. Microbiol. 37,
4099–4106
59 Russell Stuart, J. and Norvig, P. (2003) Artificial
intelligence: a modern approach, Prentice Hall
60 Quinlan, J.R. (1986) Induction of Decision
Trees. Mach. Learn. 1, 81–106
61 Rutkowski, L. (2004) Generalized regression
neural networks in time-varying environment.
IEEE Trans. Neural Netw. 15, 576–596
62 Wang, Y.M. et al. (2003) Nonlinear estimation
and modeling of fMRI data using spatio-temporal
support vector regression. Inf Process Med Imaging
18, 647–659
63 Cherkassky, V. and Ma, Y. (2004) Practical
selection of SVM parameters and noise estimation
for SVM regression. Neural Netw. 17, 113–126
64 Sreevatsan, S. et al. (1997) Ethambutol
resistance in Mycobacterium tuberculosis: critical
role of embB mutations. Antimicrob. Agents
Chemother. 41, 1677–1681
65 Kiepiela, P. et al. (2000) Genomic mutations in
the katG, inhA and aphC genes are useful for
the prediction of isoniazid resistance in
Mycobacterium tuberculosis isolates from Kwazulu
Natal, South Africa. Tuber. Lung Dis. 80, 47–56
66 Whelen, A.C. et al. (1995) Direct genotypic
detection of Mycobacterium tuberculosis rifampin
resistance in clinical specimens by using singletube heminested PCR. J. Clin. Microbiol. 33,
556–561
67 Williams, D.L. et al. (1998) Contribution of
rpoB mutations to development of rifamycin
cross-resistance in Mycobacterium tuberculosis.
Antimicrob. Agents Chemother. 42, 1853–1857
68 Flueck, T.P. et al. (2000) Correlation of in vivoresistance to chloroquine and allelic
polymorphisms in Plasmodium falciparum
isolates from Uganda. Trop. Med. Int. Health 5,
174–178
69 Nishiguchi, S. et al. (2001) Method to detect
substitutions in the interferon-sensitivitydetermining region of hepatitis C virus 1b for
prediction of response to interferon therapy.
Hepatology 33, 241–247
70 Blons, H. et al. (2004) Matrix metalloproteinase
3 polymorphism: a predictive factor of response
to neoadjuvant chemotherapy in head and
neck squamous cell carcinoma. Clin. Cancer Res.
10, 2594–2599
www.drugdiscoverytoday.com
529
Reviews • DRUG DISCOVERY TODAY: BIOSILICO
REVIEWS
DDT • Volume 10, Number 7 • April 2005